物界前沿East Voice · China AI Daily | August 4, 2026 · Issue #763
August 4, 2026 · Tuesday
I. Foundation Models / LLM
1. Alibaba Releases Qwen3.8-Max: 2.4 Trillion Parameters, Priced Below Kimi K3
Date: 2026-08-03
Summary: Alibaba unveiled its flagship MoE model Qwen3.8-Max (2.4T parameters / 95B active), with major improvements in coding and professional office tasks. API pricing at $2/M input tokens and $6/M output tokens undercuts Kimi K3’s $3/$15. Open-weight release planned for next week; supports 1M token context.
Source: Bloomberg, Nasdaq/Zacks, The Information, Alibaba official
Editor’s Take: Qwen3.8-Max isn’t just a parameter race — it’s a pricing power battle. Alibaba’s 33% cheaper API than Kimi K3 plus an open-source strategy will directly squeeze closed-source model margins.
2. Moonshot AI Denies “August IPO Filing,” But Capitalization Accelerates
Date: 2026-08-03
Summary: Reuters/IFR reported Moonshot AI may file for Hong Kong IPO as early as August to raise $3B; the company denied via insiders. However, substantive progress is clear: completed corporate restructuring in July, Series F over $3.5B closed at $35B post-money, Series G pre-money valuation set at $50B. ARR tripled from $100M in March to $300M by mid-June.
Source: Reuters, The Paper (澎湃), Toutiao
Editor’s Take: “August filing” is false, but Moonshot is already a “quasi-HK 18C candidate.” With Kimi K3 driving ARR to triple in three months and API revenue exceeding 70%, B2B monetization has begun at scale. The $50B pre-IPO valuation already surpasses Figure AI.
3. China’s “Big Four” LLM Landscape Takes Shape
Date: 2026-08-03
Summary: Comprehensive comparison of four leading Chinese models: Kimi K3 (AA Intelligence Index 57, top-ranked open-source) for validation; GLM 5.2 (MIT open-source, Code Arena #1) for coding; Qwen3.8-Max (largest official benchmark suite) for flagship; DeepSeek V4 Flash ($0.14/M) for extreme cost-efficiency. All four support 1M token context.
Source: cnblogs, CSDN
Editor’s Take: China’s foundation models have moved from “chaotic competition” to “differentiation” — flagship/cost-efficient/coding/multimodal each has its niche. Enterprises no longer need “one model to rule them all.”
4. US Tech Giants Switching to Chinese LLMs: China Models Hit 48% Token Share on OpenRouter
Date: 2026-08-03
Summary: Coinbase switched to GLM 5.2 and Kimi K2.7, cutting AI spend by nearly half; Airbnb migrated customer service from GPT to Qwen; Lindy moved from Claude to DeepSeek V4, reducing inference costs by 90%. On OpenRouter, Chinese model token share surged from under 2% a year ago to 48%, while US models dropped from 74% to 32%. Approximately 80% of US AI startups now use Chinese open-source models.
Source: Toutiao, US-China Economic and Security Review Commission Report
Editor’s Take: This isn’t a “China AI wins” narrative — it’s the market voting with its wallet. When the performance gap shrinks to 6-9 months but costs differ by 100x, the choice is rational. Political risk remains the biggest uncertainty.
5. China’s LLM “Three Musketeers” Complete Capitalization Loop
Date: 2026-08-03
Summary: Zhipu AI listed in Hong Kong (~HK$440B), Moonshot AI at $50B pre-IPO, DeepSeek at $74B first-market valuation. Together they form a complete loop across upstream compute (domestic chip day-0 adaptation), midstream models (differentiated positioning), and downstream applications (Agent/MaaS).
Source: Toutiao/Infrastructure Wind Vane
Editor’s Take: With Zhipu at HK$440B, DeepSeek at $74B, and Moonshot at $50B pre-IPO lined up, Chinese foundation models are no longer “America’s shadow” — they’re the axis around which A-share compute, HK-listed applications, and first-market agent valuations are being repriced.
6. The Guardian Reports: Chinese Open-Source Models Spark Internal US Tech Debate
Date: 2026-08-03
Summary: The Guardian reported that the rise of Chinese open-source models is creating divisions within the US tech community — one camp advocates tighter tech restrictions, while another argues open cooperation benefits innovation. Stanford’s 2026 AI Index Report concluded the US-China model performance gap has essentially disappeared.
Source: The Guardian
Editor’s Take: “The performance gap has disappeared” is a heavyweight assessment. When technology reaches parity, competition shifts from “who’s more advanced” to “who has a more open ecosystem and lower costs.”
II. AI Software / SaaS
1. A-Share Compute Sectors Cool, AI Applications Heat Up: Chuanzhi Education Hits 6-Day Streak
Date: 2026-08-04
Summary: On August 3, STAR 50 Index dropped 5.08% with semiconductors, storage, and HBM leading declines. But AI application stocks bucked the trend — Chuanzhi Education hit 6 consecutive limit-ups, Tianyu Digital and Hengyin Tech both hit 2-day streaks. Capital rotating from overvalued compute hardware to undervalued application plays.
Source: Toutiao
Editor’s Take: Classic mid-cycle style rotation — “buying shovels” gives way to “buying rent-collectors.” But Chuanzhi’s 6-day streak is sentiment-driven, not fundamental; most application stocks need 2-3 quarters to deliver earnings.
2. Desktop AI Agent Race Intensifies: “Battle of a Hundred Agents”
Date: 2026-08-03
Summary: Desktop agent products proliferating rapidly: Tencent WorkBuddy, Alibaba QoderWork/QwenWork, ByteDance Trae/Doubao, Kuaishou KroWork, ToDesk’s ToClaw. In June 2026, 17 mainstream desktop AI office agents collectively reached 60M monthly visits in China.
Source: 36Kr, Securities Times, NBD
Editor’s Take: Desktop agents are becoming the “new browser entry point” of the AI era. But severe homogeneity means the winner will be whoever first establishes differentiation through enterprise-grade security, data closed-loops, and cross-application orchestration.
3. iFlytek AI Learning Device Sales +150% YoY, WPS AI MAU Exceeds 80M
Date: 2026-08-03
Summary: AI application-layer earnings signals are clear: iFlytek AI learning device sales grew 150% YoY, Kingsoft WPS AI MAU exceeded 80M,TRS Government AI agents cover 80% of central ministries, Kunlun Tech overseas revenue reached 92% of total.
Source: Toutiao
Editor’s Take: These numbers prove AI applications aren’t “without earnings” — they were simply “overshadowed by the compute halo.” When market attention shifts from GPUs to ARR, these companies’ valuations will be rediscovered.
III. Humanoid Robots
1. Institutions Collectively Raise Humanoid Robot Shipment Forecasts: ~50K Units in China for 2026
Date: 2026-08-03
Summary: Morgan Stanley raised China’s 2026 humanoid robot shipment forecast from 28K to 50K units (44.6K by 2030, CAGR 106%). Deutsche Bank raised global forecast to ~50K (China ~40K). Goldman Sachs maintains global 51K. 2026 officially defined as “mass production year one.”
Source: Eastmoney Research Center, Toutiao
Editor’s Take: Three major investment banks simultaneously raising forecasts isn’t coincidence — it’s driven by order and capacity data. The industry narrative is shifting from “backflip demos” to “thousand-unit deliveries.”
2. Unitree Robotics Launches STAR Market IPO: Inquiry Aug 5, Subscription Aug 10
Date: 2026-07-30
Summary: Unitree Robotics disclosed its STAR Market IPO prospectus, planning to issue 40.45M shares (10% of post-IPO total), all new shares. Preliminary inquiry on August 5, subscription on August 10. H1 2026 revenue forecast: RMB 1.052-1.128B (+35-45% YoY), net profit RMB 258-306M.
Source: Unitree official filing, Gasgoo
Editor’s Take: Unitree is the first top-tier Chinese humanoid robot company pursuing A-share IPO. Revenue of RMB 1B+ and profit of RMB 258M+ provides the sector’s first hard valuation anchor.
3. AgiBot Initiates Hong Kong IPO Process: Target Valuation HK$40-50B
Date: 2026-08-03
Summary: AgiBot (智元创新) has initiated its Hong Kong listing process, potentially as early as August. Cornerstone investors target HK$40-50B valuation (~RMB 34.6-43.3B), a massive leap from 2025’s RMB 15B. As of June, 15,000 embodied intelligence robots have rolled off production lines; 2025 shipments exceeded 5,100 units; 2025 revenue surpassed RMB 1B.
Source: Gasgoo
Editor’s Take: 15,000 units produced + RMB 1B revenue + HK$40B valuation — AgiBot’s commercialization pace exceeds market expectations. Combined with Unitree, an A+H dual-listing landscape will reshape the sector’s capital pricing system.
4. Mojia Robotics Global Cumulative Deliveries Exceed 2,000 Units 
Date: 2026-07-30
Summary: Mojia Robotics announced global cumulative deliveries surpassed 2,000 units across 60+ countries and regions. Starting from Malaysian automotive marketing in 2025, expanded to Southeast Asia, Middle East, Europe, Africa, and Latin America. Secured EU CE and US FCC certifications.
Source: Gasgoo
Editor’s Take: 2,000 units × 60 countries = a scaled template for Chinese robot exports. The shift from “selling out” to “putting down roots” depends on localization services and data closed-loop capabilities.
5. Delta Intelligence Closes ~RMB 500M Angel++ Round — Sixth Funding in Six Months
Date: 2026-08-03
Summary: Delta Intelligence, incubated from Beijing Institute of General AI, closed a ~RMB 500M Angel++ round — its sixth funding round within six months of founding. Investors include AgiBot, Hillhouse Ventures, Oriza Holdings, Fosun RZ Capital, and Lenovo Capital. Funds earmarked for foundation model iteration and data closed-loop construction.
Source: Gasgoo
Editor’s Take: Six rounds in six months signals extreme capital hunger for the “robot foundation model” track. Whether foundation models translate to engineered product capability still needs time to validate.
IV. Autonomous Driving
1. Chinese Robotaxi Companies Collectively Enter Europe via Local Partnerships
Date: 2026-08-03
Summary: WeRide partnered with Denmark’s GreenMobility to enter the Nordic region (6th European market); Pony.ai partnered with Uber and Croatia’s Verne; Baidu Apollo Go road-testing with Freenow in London; Momenta secured German urban road testing permits. All using “local operator + Chinese tech” model.
Source: frontiernews.ai
Editor’s Take: Unlike Waymo’s “build and operate yourself,” Chinese firms chose “asset-light + local partnership” for overseas expansion — faster and lighter, but requiring revenue sharing. Core competitiveness will shift from market access to safety records and unit economics.
2. SiEngine Unveils 5nm “Longying-II” Cockpit-Driving Fusion Chip: 200 TOPS AI Compute
Date: 2026-08-03
Summary: According to Frost & Sullivan, SiEngine ranked #1 in China’s cockpit domain controller SoC installations in 2025. Its 5nm “Longying-II” cockpit-driving fusion chip delivers 200 TOPS AI compute, with adaptation planned for Q1 2027.
Source: Gasgoo
Editor’s Take: Cockpit-driving fusion is the next battlefield for automotive chips. 200 TOPS of AI compute means cockpit and autonomous driving can share a single chip’s resources, dramatically reducing BOM costs.
V. World Models / Physical AI
1. World Models Officially Transition from Academic Concept to Engineering Deployment
Date: 2026-08-04
Summary: In 2026, multiple AI teams released models and data collection solutions targeting physical world perception, spanning retail, hospitality, and outdoor scenarios. World models are officially moving from papers to products, reaching a commercialization inflection point.
Source: Weihezhisuan
Editor’s Take: When world models start appearing in retail and hospitality solution packages, it signals the technology has crossed from “demo phase” to “deliverable phase.”
2. Six Structural Shifts in Digital Twin Industry: AI Becomes Core Watershed
Date: 2026-08-04
Summary: H2 2026 digital twin industry faces six transformations: ①Vertical scenarios replace general platforms; ②Visualization upgrades to intelligence, AI becomes watershed; ③Business models shift from project delivery to long-term operations; ④L3 simulation and L4 intelligence segments growing fastest; ⑤Vendors without AI capabilities will be eliminated by 2027; ⑥Data closed-loops become core competitiveness.
Source: Industrial Internet Wind Vane
Editor’s Take: “Vendors without AI capabilities will be eliminated by 2027” isn’t alarmist — when clients discover your digital twin is just a “pretty dashboard” that can’t run decision simulations, contracts won’t renew.
3. ACE Robotics Open-Sources Kairos 3.0-4B: Native Embodied Intelligence World Model
Date: 2026-08-04
Summary: ACE Robotics (subsidiary of StepFun) open-sourced Kairos 3.0-4B, achieving physics-level deep understanding and cross-embodiment generalization with just 4B parameters. Achieves 1:1.5 real-time edge generation on NVIDIA THOR platform, with inference speed 72x faster than Cosmos 2.5. 7-minute coherent interaction videos set new industry benchmark.
Source: The Times Australia
Editor’s Take: 4B parameters outperforming hundred-billion-parameter models — the key is “native embodied design.” Not bolting motion interfaces onto a language model, but building world understanding from physical laws. 72x inference speed makes edge deployment reality.
4. Humanoid Robot Industry Enters “Commercial Delivery Phase”
Date: 2026-08-03
Summary: The humanoid robot industry underwent three major shifts in 2026: ①Competition moved from athletic performance demos to long-duration task execution; ②Business models shifted from selling equipment to “Robot-as-a-Service” (RaaS); ③Supply chain focus migrated from robot bodies to core components and data. High-quality industrial scene data has become the critical bottleneck for training VLA models.
Source: Toutiao/Eastmoney Research Center
Editor’s Take: Whoever controls more real production-line data gains the advantage in model iteration — this is the core competitive moat in the humanoid robot track, and the biggest defensive advantage for mass-production leaders like Unitree, AgiBot, and BYD.
东声 East Voice — Physix Frontier’s China AI column for global readers
东声西达，中国视角。
© 2026 Physix Frontier. All rights reserved.